Вы находитесь на странице: 1из 15

ARTICLE IN PRESS

Mechanical Systems
and
Signal Processing
Mechanical Systems and Signal Processing 18 (2004) 833847
www.elsevier.com/locate/jnlabr/ymssp

Utilising statistical residual life estimates of bearings to


quantify the inuence of preventive maintenance actions
Pieter-Jan Vlok, Maciej Wnek*, Maciej Zygmunt
ABB Corporate Research Center, ul. Starowislna 13A, 31-038 Krakow, Poland
Received 11 January 2003; received in revised form 6 June 2003; accepted 15 September 2003

Abstract

In this paper it is illustrated how statistical residual life estimates of bearings can be used to justify
maintenance practices. Residual life estimates are based on Proportional Intensity Models for non-
repairable systems utilising historic failure data and corresponding diagnostic measurements. A case study
is presented where failure and diagnostic data obtained from roller bearings operating in the dryer section
of a paper machine are used to predict future failure times of bearings. If these predictions are compared to
the diagnostic measurements, i.e. vibration and lubrication levels, it becomes evident how changes in these
diagnostic measurements inuence the residual life of the bearings. From this it is possible to justify
maintenance practices.
r 2003 Elsevier Ltd. All rights reserved.

Keywords: Lifetime estimate; Bearing; Maintenance quality

1. Introduction

Preventive maintenance is often initiated on industrial equipment based on observed diagnostic


measurements such as vibration levels, temperatures, pressures, etc., that are outside of certain
acceptable levels. After performing the preventive maintenance the success of this action is judged
on a binary basis. If the said diagnostic measurements have decreased or increased sufciently to
be within specication, the preventive maintenance action is considered to be successful. If not,
additional actions are often instigated to get all diagnostic measurements within specication.
In this paper we describe a methodology to quantify the success of preventive maintenance
more accurately than a simple binary number. The residual life estimation methodology of Vlok
[1] is applied but not with the sole purpose to predict future failures. It is shown how these
estimates can be used to quantify the success and efciency of preventive maintenance actions.

*Corresponding author. Fax: +48-12-4244101.


E-mail address: maciej.wnek@pl.abb.com (M. Wnek).

0888-3270/$ - see front matter r 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ymssp.2003.09.003
ARTICLE IN PRESS

834 P.-J. Vlok et al. / Mechanical Systems and Signal Processing 18 (2004) 833847

The paper starts off with an introduction to the residual life estimation concept and describes
the data required to perform the calculations. It further discusses how this concept can be used to
quantify the success of preventive maintenance actions. The validity of the theory is illustrated
with a case study in the last section.

2. Residual life estimation of equipment based on failure intensity proportions

In recent research a methodology suitable to estimate the residual life of industrial equipment
was developed based on actuarial statistics (see [1,2]). Life insurance companies need to make an
estimate on a clients residual life at his/her current age to be able to set appropriate insurance
premiums. These companies compile data of the ages of people (with certain similarities) at death
and their associated life styles, e.g. the number of cigarettes smoked per day, number of beers per
week, stress levels, etc. This data is represented by a Proportional Intensity Model (PIM) and it is
hence possible to estimate the residual life of a living client at a certain age with a certain life style.
The methodology for machines makes use of the ages of machines at failure and associated
diagnostic information that was collected during the machines lifetime. The concept is shown in
Fig. 1. Data of similar failed machines is compiled and a PIM specially developed for reliability
data is tted to observations. From here it is possible to estimate the residual life of a machine
currently in operation, at a certain age and with certain diagnostic measurements. The
methodology, its terminology, data requirements and outputs are described in the remainder of
this paper.

2.1. Terminology and notation

The terminology and notation of Ascher and Feingold [3] is used in this paper for clarity. First,
proper terminology for different types of machines and their components is required:
(i) Part. An item that is never disassembled and is discarded after its rst failure, i.e. the rst
time it ceases to perform satisfactory.

Fig. 1. Global overview of the RLE concept.


ARTICLE IN PRESS

P.-J. Vlok et al. / Mechanical Systems and Signal Processing 18 (2004) 833847 835

(ii) Socket. A space that, at any given time, holds a part of a given type.
(iii) System. A collection of two or more sockets with their associated parts that is interconnected
to perform a specic function. All the associated parts of a system do not necessarily need to
perform satisfactory for the system to perform satisfactory.
(iv) Non-repairable system. A system that is discarded after its rst failure on system-level.
(v) Repairable system. A system that can be restored to perform its function by any method other
than complete replacement after its rst failure on system-level.

These denitions are very important because the residual life of non-repairable systems are
modelled with fundamentally different statistical techniques than repairable systems. Systems are
primarily categorised as either non-repairable or repairable based on the properties of the failure
data that it generates and secondarily by its physical properties. A system that produces
independent and identically distributed failure data is categorised as non-repairable while a
system that generates failure data with an increasing or decreasing trend is dened as repairable.
A gearbox, for example, is from a physical point of view a repairable system but in the early stages
of its operation it could produce independent and identically distributed failure times and it will
be classied as non-repairable in residual life analyses. Differences between non-repairable and
repairable systems are discussed in detail in sections to follow.
The denition of system-failure should be formulated clearly for every particular situation since
the objective of residual life analysis is to simulate the occurrence of failures. In most situations in
practice systems are not allowed to age undisturbed, i.e. preventive maintenance is performed, and
this should be accounted for when constructing lifetime models. There are three ways to deal with
the effects of preventive maintenance on failure times:
(i) Assume that preventive maintenance on a particular system or a group of similar systems has
always got the same effect on the lifetime regardless of the frequency or the type of preventive
maintenance. Also assume that the effect of preventive maintenance will remain constant for
future systems lifetimes. This approach should preferably be avoided since it basically
ignores preventive maintenance and would only be followed when information about the
occurrences or types of preventive maintenance is unavailable.
(ii) Include the times of preventive maintenance as truncated failure time observations (also
called suspensions) in a data set. Suspensions are considered as partial information by
regression coefcients of reliability models, i.e. it is only known that the system survived up
to the age of preventive maintenance but does not provide any information as to when the
system would have failed if left undisturbed.
(iii) Assume that diagnostic information that describes the condition of a system or the conditions
under which a system operates for example vibration levels, temperatures, pressures, loads,
etc., also quanties the effects of preventive maintenance. Include these parameters in lifetime
models to better describe a systems recorded and future survival times. This approach is
considerably more complex than the rst two but also generally yields by far the best results.
This paper focuses on the third approach.
Diagnostic measurements recorded during a systems lifetime are referred to as covariates in
residual life analysis. Covariates could be dependent or independent of time and any quantity or
combination of quantities that describe a systems condition can be included in reliability models.
ARTICLE IN PRESS

836 P.-J. Vlok et al. / Mechanical Systems and Signal Processing 18 (2004) 833847

The number of covariates that are allowed in a reliability model is limited by the number of
failures in a data set. More failures allow for more covariates. Covariates that are measured on a
system are denoted by zk1 ; zk2 ; y; zkm ; where m is the total number of covariates and k is the system
number (assuming several similar systems are considered in an analysis).
Before residual life models for non-repairable and repairable systems are introduced, it is
necessary to dene appropriate time scales to measure lifetimes. Consider a particular system k
that is part of a group of w similar machines on which one covariate is measured. An example
sample path of the failure process of system k is shown in Fig. 2. Assume that the sudden
decreases in covariate levels are due to preventive maintenance that was performed.
Xik ; i 1; 2; 3; y; in Fig. 2 refers to the interarrival time between the (i1)th failure and ith
failure of system k. Xik is a random variable with X0k  0: This is referred to as local time and is
convenient to use when analysing non-repairable systems. The real variable xki measures the time
elapsed since the most recent event. An auxiliary variable Cik is also recorded with each Xik to
indicate whether Xik was a real failure or a truncated failure observation, i.e. a suspension.
Cik 0 in the case of a suspension and Cik 1 in the case of a failure.
Tik ,i 1; 2; 3; y; measures time from 0 to the ith failure time. Tik is also called the arrival time to
the ith failure and is mostly used to analyse repairable systems. This time scale is referred to as
global time. It follows that Trk X1 X2 ?Xr ; where r is total number of observed failures. As
before, an auxiliary variable Cik is also recorded with each Tik .
From the time scales, a random variable N k t is dened as the maximum value of r for which
Tr ptk ; i.e. N k t is the number of failures that have occurred in 0; tk : N k t; tk X0 is the integer
k

valued counting process that includes both the number of failures in 0; tk ; N k t; and the instants
of occurrence, T1k ; T2k ; y:
Data collection is often stopped in-between failures to perform an updated reliability analysis
with all data available up to the particular point. In such cases the instant at which data collection
was stopped should be included in the data set as a temporary suspension. Temporary suspensions
are considered in the same manner as normal suspensions in reliability models and the time of the
temporary suspension should be recorded in Xik or Tik with Cik 0:

Fig. 2. Example sample path of a failure process.


ARTICLE IN PRESS

P.-J. Vlok et al. / Mechanical Systems and Signal Processing 18 (2004) 833847 837

2.2. Data requirements

The quality and usefulness of reliability models are directly proportional to the accuracy of the
data that it is based upon. It is recommended to collect data in dedicated tables to avoid
unnecessary calculation errors while following these steps:
(i) Identify the particular situation of a system or group of systems that residual life analysis will
be performed on. Fig. 3 illustrates four situations as a function of preventive maintenance
and the availability of diagnostic information. This paper focuses on the top-left situation,
i.e. where preventive maintenance is performed and diagnostic data is available.
(ii) Select a convenient time scale, i.e. local time or global time, and be consistent with it. Local
time can be converted to global time and vice versa after a system or group of systems have
been classied as non-repairable or repairable.
(iii) Generate an events table containing the failure times of each system with its corresponding
failure type indicator.
(iv) Generate an inspections table containing records of the measured covariates and the
corresponding times at which measurements were done.
Proper structuring of a data set can scarcely be overemphasised. The majority of calculation
errors originate from poorly structured data sets.

2.3. Methodology

Certain properties of a data set rst need to be evaluated to be able to select the most
appropriate type of lifetime model for the data set. Fig. 4 shows an outline of the necessary steps if
only one system is considered. First it is required to test for a trend in the chronologically ordered
data set, i.e. a tendency of interarrival times to increase or decrease. Laplaces (see [46]) trend test
is most often used for this purpose. If a trend is present (increasing or decreasing), a lifetime
model suitable for repairable systems should be selected for the data set. If no trend is present, it is
required to test for dependencies between interarrival times. This step is most often omitted since

Fig. 3. Data collection from different situations.


ARTICLE IN PRESS

838 P.-J. Vlok et al. / Mechanical Systems and Signal Processing 18 (2004) 833847

Chronologically
ordered Xi's

Yes Repairable systems


Trend?
theory

No

Xi's identically distributed but not


necessarily dependent

Yes Branching Poisson


Dependence?
Process

No

Non-repairable
systems theory

Fig. 4. Statistical failure analysis of successive interarrival times of a single system (Adapted from Ascher and Feingold,
1984).

approximately 30 failure observations are required to perform such a test with condence
according to Cox and Lewis [7]. It is hence assumed that there is no dependence between
interarrival times and non-repairable systems theory should be used for data sets of this type.
Most often a group of similar systems is considered because of data scarcity. Similarities amongst
systems are accommodated in the residual life estimation methodology through a process called
stratification. Suppose a group of k 1; 2; y; w similar systems are considered and qk events have
been observed on each of the w systems. If the residual life of a particular system k wants to be
estimated the following steps should be taken:
(i) Find all systems with N k tX4:
(ii) Determine which of these systems event times display a trend.
(iii) Renumber the systems such that j 1; 2; y; m labels the systems without a trend and
i 1; 2; y; n labels the systems with a trend.
From this several options are available:
P
(i) The case where kAj and m j
j1 N tX8
(a) Model the m systems with non-repairable systems theory and predict the arrival time of
N k t 1:
(b) Stratication
P j amongst any combination of the m systems is permissible provided that
j N tX8 for the selected systems.
(c) If there are valid physical reasons toPinclude some or all of the w m n systems as non-
repairable systems in order to get m j
j1 N tX8 it is allowed but should preferably be
avoided.
(d) The n systems that have been classied as repairable systems should under no
circumstances be stratied with any of the m systems.
ARTICLE IN PRESS

P.-J. Vlok et al. / Mechanical Systems and Signal Processing 18 (2004) 833847 839
P
(ii) The case where kAi and ni1 N i tX8
(a) Model the n systems with repairable systems theory and predict the arrival time of
N k t 1:
(b) Stratication
P i amongst any combination of the n systems is permissible provided that
i N tX8 for the selected systems.
(c) If there are valid physical reasons toPinclude some or all of the w  m  n systems as non-
repairable systems in order to get ni1 N i tX8 it is allowed but should preferably be
avoided.
(d) The m systems that have been classied as repairable systems should under no
circumstances be stratied with any of the n systems.

P Pn
(iii) The case where m j
j1 N tX8 or
i k
i1 N tX8 but N to4
(a) If there are valid physical reasons to include system k with either the m non-repairable or
the n repairable systems, it is allowed but should be done with great care.
(b) The risk of unrealistic estimations is very high when following this option.

(iv) The case where wX8 and N k t > N k t for all k


(a) Consider all the N k tth lifetimes of the w systems as failure data obtained from a virtual
non-repairable system and predict the arrival time of N k t 1.
(b) It should be emphasised that this step is only allowed if w X 8.

These steps are discussed in detail in sections to follow.

2.4. Residual life model for non-repairable systems

For non-repairable systems, a statistical property called the Force of Mortality (FOM) is
modelled in a PIM as a basis to calculate residual life form. A systems FOM is its instantaneous
probability of failure at any given time, x, provided that the system has survived up to x. Most
mechanical systems have increasing (not necessarily a monotonically increasing) FOMs, i.e. the
instantaneous probability of failure increases as time progresses. Electronic equipment often
exhibit decreasing FOMs which corresponds to the well known infant mortality concept.
The FOM is often regarded as the most useful entity in statistical failure analysis because it
provides an early indication of the most appropriate maintenance action. For a FOM increasing
with x a non-repairable system has an increasing probability to fail and use-based preventive
action is a denite option to consider, although costs dictate the nal decision. Preventive action
will only be used if the total cost of failure is considerably higher than the total cost of prevention.
If the FOM remains roughly constant with x the non-repairable system has a random shock
failure pattern because the instantaneous risk of failure of the non-repairable system remains
constant throughout its life. Corrective action is the rst option to consider for this case (if
condition monitoring techniques are not feasible). Corrective renewal will also most probably be
used for a non-repairable system with a decreasing FOM since the probability of failure reduces as
time progresses. Condition-based maintenance, such as vibration monitoring, could be used for
any shape of the FOM.
ARTICLE IN PRESS

840 P.-J. Vlok et al. / Mechanical Systems and Signal Processing 18 (2004) 833847

It is proposed that the FOM of any system, k, under consideration at any point in time, x, and
for any stratum, s, is represented by
!bks 1 P m P m
gksj zki aksj zki
bks cks x  tks j j
hx; h k k
e j1
e j1
; 1
Zs Zs
where cks ;tks and the two exponential terms allows for accelerations, decelerations, setbacks,
advances and discontinuities in the FOM caused by operating conditions and preventive
maintenance. bks ; Zks ; gksj and aksj are all regression coefcients. For convenience, h is dened as a
vector containing all of the models unknowns. If a particular situation requires it or data is
limited, it is possible to make simplications to (1) by, for example, restricting coefcients to be
stratum specic but not system-copy specic, thereby reducing the total number of regression
coefcients. For a comprehensive discussion on the proposed model, more general models and
model tting techniques, see (1).
When hx; h is known, it is possible to calculate the probability density, f x; h; of failures for
the system by
 Z x 
f x; h hx; h exp  hs; h ds 2
0

and hence the expected value of f x; h by


RN
s f s; h ds
EXr1  Rx N : 3
x f s; h ds
The residual life until the r 1th failure is mr1 EXr1   x: The calculation for the upper
and lower bounds around the estimate are calculated in the same manner.
For the calculation of (2) it is required to extrapolate measured covariate values. Appropriate
techniques for the extrapolation are described by Vlok [1] and Makis and Jardine [8].

2.5. Residual life model for repairable systems

For repairable systems, a statistical property called the rate of occurrence of failure (ROCOF)
is modelled in a PIM as a basis to calculate residual life form. As the name suggests, is the
ROCOF simply the time derivative of the expected number of failures up to any instant t. This is
strictly speaking only correct if a non-homogeneous Poisson process (NHPP) is used to model the
ROCOF but since a NHPP will be used in this paper no further comments are made. The ROCOF
is a monotonically increasing function since the number of failures recorded on a system cannot
decrease.
It is proposed that the ROCOF of any system, k, under consideration at any point in time, t,
and for any stratum, s, is represented by
P m
k k
Pm
k k
k
gs zi as zi
k k k j j j j
rt; h eGs cs Ys tts e j1 e j1 ; 4
where cks ;tks and the last two exponential terms allows for accelerations, decelerations, setbacks,
advances and discontinuities in the ROCOF caused by operating conditions and preventive
maintenance. Gks ;Uks ; gksj and aksj are all regression coefcients. For convenience, h is dened as a
ARTICLE IN PRESS

P.-J. Vlok et al. / Mechanical Systems and Signal Processing 18 (2004) 833847 841

vector containing all of the models unknowns. If a particular situation requires it or data is
limited, it is possible to make simplications to (4) by, for example, restricting coefcients to be
stratum specic but not system-copy specic. For a comprehensive discussion on the proposed
model, more general models and model tting techniques, see [1].
When rt; h is known, it is possible to calculate the expected time to the r 1th, i.e. ETr1 ;
by solving numerically for ETr1  in
Z ETr1  Z t
rs; h ds 1  rs; h ds; 5
t Ti

where covariate extrapolation should be done in the same manner as above. The residual life of
the system is thus mr1 ETr1   t.
Calculation of condence bounds around the estimate is more complicated than for the non-
repairable case. It is required to calculate the Fischer information matrix for rt; h which involves
taking rst and second partial derivatives of rt; h with respect to time as well as with respect to
the regression coefcients. This is a fairly complicated task and will not be described in this paper.

3. Case study

A large paper mill has had numerous failures on the felt bearings of their paper machines over
the past 2 years. Most of these failures were considered as premature compared to industry
standards. As part of an effort to reduce the number of failures and extend the interarrival times
of failure, the bearing supplier was changed and lubrication levels on these bearings were reduced
on two occasions. These steps appear to be very effective and felt bearings are running
considerably longer ever since.
Despite successfully elongating the felt bearings lives, it was still of interest to the mills
management team to predict the time of future failures on these bearings and to quantify the effect
of their preventive maintenance actions.
Compiling the necessary data was based on accurate failure records led by the senior
mechanical engineer at the plant and ABBs on-line vibration condition monitoring system, Smart
Advisor, that has been in use for more than two years.

3.1. Summary of collected data

Fig. 5 shows a global overview of the interarrival times to failure of 12 of the felt bearings. It is
clear from a simple eye-ball analysis that bearings failed much more frequently during the period
of November 2000September 2001, compared to October 2001October 2002.
To explain the arrival times of the observed failures and also to predict future failures, vibration
measurements were extracted from the Smart Advisor database as well as lubrication records
from the responsible engineer. Table 1 presents a summary of the explanatory information that
was extracted.
Where BPFI, BPFO, BSF denote vibration velocities at the given frequency, OV denotes RMS
of the vibration velocity over the whole spectra and the corresponding high frequency envelope
(HFE) give the HFE amplitude at the given frequencies, respectively. Vibration data was collected
ARTICLE IN PRESS

842 P.-J. Vlok et al. / Mechanical Systems and Signal Processing 18 (2004) 833847

Fig. 5. Global overview of bearing failure times (Dots denote failures).

Table 1
Summary of covariates
Abbreviation Description
1. BPFI Ball pass frequency inner race
2. BPFO Ball pass frequency outer race
3. BSF Ball spin frequency
4. OV Overall spectral energy
5. HFEBPFI High frequency envelope ball pass frequency inner race
6. HFEBPFO High frequency envelope ball pass frequency outer race
7. HFEBSF High frequency envelope ball spin frequency
8. HFEOV High frequency envelope overall spectral energy
9. GREASE Quantity (grams) of grease applied to bearings per week

at rotating speed about 530 rpm with maximum frequency 500 Hz from velocity spectrum of 800
lines resolution ( 4 averages, Haning window). It is known that envelope spectra is more efcient
in bearing failure detection but that type of data had been collected just for last 6 months at the
plant whereas we needed about 1.5 year of history to achieve right statistical pool of events (see
for instance [9]). We were not able to inuence other measurements set-ups since we had to
conform with the plant preferred system conguration. Apart from greasing and replacement
dates no other diagnostic data was available. We were thus forced to use traditional vibration
velocities for further calculations. All the bearings taken into consideration were physically
identical and provided by the same manufacturer.
ARTICLE IN PRESS

P.-J. Vlok et al. / Mechanical Systems and Signal Processing 18 (2004) 833847 843

3.2. Construction of the model

The methodology described in Section 2.3 was used as a roadmap to the best possible model.
Following the tests for trend, it was decided that Option (i) of Section 2.3 would be the most
appropriate because of the limited number of events on any particular system.
Only bearings D211B, D211F, D212B, D242B, D243B and D242F (see Fig. 5) were used to
construct the lifetime model since these bearings have all experienced at least one failure. It is
possible to include bearings that have not failed yet in the model but this could lead to biased
regression coefcients.
It is assumed that the model determined by the data from the above bearings will also be
applicable to all the other bearings. This assumption is allowed because of the stratication
abilities of the lifetime model for non-repairable systems.
Closer inspection of the data revealed that measurements of the HFE covariates only started
fairly recently and if these covariates were included in the analysis it would decrease the data set
size signicantly. It was hence decided to omit the HFE information. Numerous tests for
statistical signicance (see [1012]) showed that BPFI, BPFO and GREASE describe the
occurrence of the observed failures in the data set best.
In the nal lifetime model all the regression coefcients were positive which means that an
increase in any of the covariates would mean a decrease in residual lifealso for the grease
amount.

3.3. Evaluation of the model

The prediction results on two bearings are shown below to illustrate the validity of the
approach: Bearing D211B that was part of the data used to t the model and Bearing D240B that
was not included in the data used to t the model but the model derived has been applied to the
Bearing D240B diagnostic data allowing for calculation lifetime estimates as well. In both cases
the residual lifetime estimates and their condence bounds are plotted at the same time as the
diagnostic measurements.

3.3.1. Bearing D211B


Fig. 6 shows the prediction history of Bearing D211B within 90% condence bands. It is clear
that on two occasions something signicantly improved the bearings life, i.e. at 175 days and
again at 400 days. At both these instants the residual life estimates increased by at least 100%. In
the period between 280400 days no measurements were taken on the bearing. Following this
observation, the residual life estimates were plotted together with the covariates. See Fig. 7.
Fig. 7 shows that on both occasions when the lubrication quantities were reduced, the residual
life estimates increased. This is somewhat different than expected but an investigation by the
bearing manufacturer conrmed the validity of these observations. The bearings were overgreased
and generated excessive heat in the processfrom there the large number of premature failures.
One can observe also correlations between vibrations levels (BPFI and BPFO) and the lifetime
prediction valuesespecially at the end a huge decrease of the lifetime value correlated with very
high vibration velocity values. Since we were not able to acquire detailed process history we are
not able to combine that with events on the production line. Similarly with the periodical
ARTICLE IN PRESS

844 P.-J. Vlok et al. / Mechanical Systems and Signal Processing 18 (2004) 833847

350

300

250
Residual life (days)

200

150

100

50

0
100 150 200 250 300 350 400 450 500
Global time (days)

Fig. 6. Prediction history of D211b (Lifetime 3).

0.7 350

Residual life estimate


GREASE
0.6 300
BPFI
BPFO

0.5 250

Residual life (days)/GREASE (gr/week)


0.4 200
Vibration (mm/s)

0.3 150

0.2 100

0.1 50

0 0
100 150 200 250 300 350 400 450 500
Global age (days)

Fig. 7. Prediction history of D211b (Lifetime 3) with associated covariates.

oscillation of RLE estimates (about 2 weeks period). This can be related to the time of grease
injection, product changes or speed changes (once every two weeks for 2 days). The detailed study
is however out of the scope of this paper. One can also observe measured correlations between
ARTICLE IN PRESS

P.-J. Vlok et al. / Mechanical Systems and Signal Processing 18 (2004) 833847 845

300

250

200
Residual life (days)

150

100

50

0
-100 0 100 200 300 400 500 600 700
Global time (days)

Fig. 8. Prediction history of D240b (Lifetime 1) with associated covariates.

greasing level end vibration covariateslowering grease amount increases slightly vibration levels
of the system.

3.3.2. Bearing D240b


Fig. 8 shows the prediction history of Bearing D240B within 90% condence bands. Contrary
to Bearing D211b it shows only one signicant improvement in the bearings condition, i.e. at
around 300 days. As before, the residual life estimates were plotted together with the covariates.
See Fig. 9.
Fig. 9 shows that the rst decrease in lubrication at around 300 days caused the increase in the
residual life estimates. When the amount of grease was reduced for the second time at 470 days, it
did not inuence the residual life estimates signicantly. Closer inspection of the graph reveals
that when the amount of grease was reduced, the two vibrations covariates increased signicantly,
thereby nullifying any positive effects on the residual life estimates. It seems thus that this bearing
was also overgreased but the second reduction of the amount of grease was too much and it
caused an increased vibration level. Since there were no major changes in the process itselflike
paper grade or composition, bigger refurbishment works, etc. we may link these two phenomena
together.

4. Conclusion

As stated in Section 3.1 (see Fig. 5) one can often observe some general patterns in the failure
statistics and at the same time introduce some policy changes (in our case it is the amount of
ARTICLE IN PRESS

846 P.-J. Vlok et al. / Mechanical Systems and Signal Processing 18 (2004) 833847

0.7 350
Residual life estimate
GREASE
BPFI
0.6 300
BPFO

0.5 250

Residual life (days)/GREASE (gr/week)


0.4 200
Vibration (mm/s)

0.3 150

0.2 100

0.1 50

0 0
0 100 200 300 400 500 600 700
Global age (days)

Fig. 9. Prediction history of D240b (Lifetime 1).

grease)but are we able to assign any procedural changes as responsible for lower failures rate?
And, even more important, can we reduce the number of failures even further by introducing
more changes to the maintenance procedure?
We tried to answer these questions and provide guidance to quantitative judgement of
maintenance procedures. Our simple case based on lubrication procedures illustrates well how the
residual life estimation can be utilised as a benchmarking parameter. RLE values allow for direct
comparison of the health and performance of machinery.
It is obvious that one can study several maintenance tasks at the same timeintroducing them
as covariates and observe the changes they impose on the resulting RLE values. A requirement for
this enough statistical evidence from the past to allow the model to converge. It is attractive that
when new maintenance policies/recipies are introduced, it is possible to judge their inuence
reasonably fast, i.e. after only 23 diagnostic measurements.
Standard maintenance procedures are often prescribed by the vendors for general operating
conditions but are not valid when equipment are subjected to very different environments, e.g.
high temperatures, high humidity, corrosive substances, etc. The RLE approach allows to
individualise the policies for each piece of equipmenteven if technically identical. This statement
is illustrated by the case of the two bearings. For Bearing D211b, the second reduction in greasing
still resulted in extension of the RLE, i.e. it had positive inuenceit is not obvious for Bearing
D240b. Here the increase in vibration levels nullied the positive impact resulting from change of
the greasing level. The reason is that these bearings operate in different sections on the paper
machine under different loads and temperatures.
This study illustrate just possibilities given by statistical lifetime estimate as a benchmark tool
since we have to comply with historical data and records of the plant we did not control the whole
ARTICLE IN PRESS

P.-J. Vlok et al. / Mechanical Systems and Signal Processing 18 (2004) 833847 847

experimental space. While using RLE as a benchmarking tool one should introduce procedures of
monitoring process changes that could impact additionally measured covariates. But since in our
case we observed similar behaviour related to greasing change for all the bearings used for
analysis [6] and knowing in addition that process did not change one may thus exclude some
uncontrolled phenomena that could change the results.
This study shows that RLE is a very convenient tool to justify, benchmark and optimise
maintenance practices.

References

[1] P.J. Vlok, Dynamic residual life estimation of industrial equipment based on failure intensity proportions, Ph.D.
Thesis, University of Pretoria, 2001.
[2] P.J. Vlok, S.J. Claasen, Estimating residual life of industrial equipment using combined proportional intensity
models, Journal of the Operational Research Society, 2002, submitted for publication.
[3] H.E. Ascher, H. Feingold, Repairable Systems Reliability: Modeling Inference, Misconceptions and Their Causes,
Marcel Dekker, New York, 1984.
[4] P.S. De Laplace, Memoire sur la probabilite des causes par les evenemens, Memoires de Mathematique et de
Physique 6 (1974) 621656.
[5] D.J. Bartholomew, Discussion of Cox (1955), Journal of the Royal Statistical Society 17 (1955) 162163.
[6] D.R. Cox, Some statistical methods connected with series of events, Journal of the Royal Statistical Society B 17
(1955) 129164.
[7] D.R. Cox, P.A. Lewis, The Statistical Analysis of Series of Events, Metheun, London, 1966.
[8] V. Makis, A.K.S. Jardine, Computation of optimal policies in replacement models, IMA Journal of Mathematics
Applied in Business Industry 3 (1992) 169175.
[9] R.B. Randall, in: C.M. Harris (Ed.), Spectrum Analysers and their Use, Shock and Vibration Handbook, 4th
Edition, McGraw-Hill, New York, 1955.
[10] D.W. Hosmer, S. Lemeshow, Applied Survival Analysis, Wiley, New York, 1999.
[11] Y. Sakamoto, M. Ishiguro, G. Kitagawa, AKAKIE Information Criterion Statistics, Reidel, Boston, 1983.
[12] M.J. Crowder, A.C. Kimber, R.L. Smith, T.J. Sweeting, Statistical Analysis of Reliability Data, Chapman and
Hall, London, 1991.

Вам также может понравиться